[Devel] Re: [PATCH 2/4] signals: send_signal: use si_fromuser() to detect from_ancestor_ns

[Oleg Nesterov]

On 10/05, Sukadev Bhattiprolu wrote:
>
> Oleg Nesterov [oleg at redhat.com] wrote:
> |
> | --- TTT_32/kernel/signal.c~FU_2_SEND_SIGNAL	2009-10-04 02:21:55.000000000 +0200
> | +++ TTT_32/kernel/signal.c	2009-10-04 03:09:44.000000000 +0200
> | @@ -928,9 +928,8 @@ static int send_signal(int sig, struct s
> |  	int from_ancestor_ns = 0;
> |
> |  #ifdef CONFIG_PID_NS
> | -	if (!is_si_special(info) && SI_FROMUSER(info) &&
> | -			task_pid_nr_ns(current, task_active_pid_ns(t)) <= 0)
> | -		from_ancestor_ns = 1;
> | +	from_ancestor_ns = si_fromuser(info) &&
> | +			   !task_pid_nr_ns(current, task_active_pid_ns(t));
>
> Makes sense. And we had mentioned earlier that container-init is immune
> to suicide but should we add a check for 'current == t' above to cover the
>
> 	send_sig(SIGKILL, current, 0);
>
> in load_aout_binary() and friends
>
> 	from_ancestor_ns = si_fromuser(info) && (current == t ||
> 				!task_pid_nr_ns(current, task_active_pid_ns(t)));

I don't think so.

First of all, this is just ugly. If we need this check we should change
the callers, not send_signal().

But more importantly, I disagree with "container-init is immune to suicide"
above. This is another issue I was going to discuss later, lets do this now.

When load_elf_binary() does send_sig(SIGKILL, current) init must die, because
we have no option. Exec failed, but we can't return to user-space with the
error code, it is too late.

So, imho this patch also fixes this case by accident, but I think it would
be better to change load_aout_binary/etc to use force_sig_info() to make
the code more explicit.

What do you think?

Oleg.

_______________________________________________
Containers mailing list
Containers at lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers